A method for identifying languages from positive data alone
A trained scorer can rank most positive examples above out-of-language perturbations.
A trained scorer achieves PAC conditional identification from positive data alone. It ranks all but an -fraction of positives above out-of-language perturbations in their own orbit.
The process uses an innate recursive contrast operator . This operator maps any string to a finite set of structurally bounded perturbations of it.
A discriminative scorer of bounded capacity is trained on self-generated perturbations. This training uses InfoNCE-style contrastive estimation.
The negatives are constructed by the learner, not supplied by an oracle. No global normalizing constant over strings is ever computed.
The misranking bound follows from the geometry of the contrastive loss. This includes per-example loss floors and a definite loss increment under misranking.
The work instantiates hypotheses for substitutable context-free languages and subregular classes. An open conjecture seeks to locate the construction in the mildly context-sensitive hierarchy.
Key facts
- The scorer achieves PAC conditional identification from positive data alone.
- Negatives are constructed by the learner and not supplied by an oracle.
- The misranking bound follows from the geometry of the contrastive loss.